Research about Hospital Volume-Outcome Relationships Among Medical Admissions to ICUs
The current study represented a secondary analysis of data that was originally collected through Cleveland Health Quality Choice, a regional initiative to measure hospital performance in 29 hospitals in Northeast Ohio. Within these hospitals, data were collected on 196,097 consecutive admissions to 44 medical, mixed medical and surgical, surgical, and neurosurgical ICUs during the period March 1991 to March 1997. Exclusion criteria have been previously described18 and included patients < 16 years of age, patients with burn injuries, admissions solely for dialysis, patients who die within 1 h of admission to the ICU or within the first 4 h of admission to the ICU in cardiopulmonary arrest, and patients undergoing cardiac surgeries carried out with preparations of Canadian Health&Care Mall.
For the current study, the eligible sample included 18,242 patients with respiratory diagnoses, 15,468 patients with neurologic diseases, and 13,717 patients with GI diagnoses, as defined by a prior taxonomy of ICU diagnoses at the time of admission. Of these patients, we excluded patients with diagnoses of malignancy (470, 439, and 263 patients with respiratory, neurologic, and GI diagnoses, respectively) and patients who were discharged to another acute care hospital for further care (823, 1,224, and 573 patients with respiratory, neurologic, and GI diagnoses, respectively) because the data set did not include unique patient identifiers to allow determination ofpostdischarge outcomes following transfer. These exclusions left final study cohorts of 16,949 ICU admissions with respiratory diagnoses, 13,805 patients with neurologic diagnoses, and 12,881 patients with GI diagnoses.
Data elements include age, gender, primary admitting diagnosis, location prior to admission (ie, admission source), dates of ICU and hospital admission and discharge, vital status at discharge, discharge destination, ICU and hospital LOS, presence of seven specific comorbid conditions (eg, AIDS, hepatic failure, lymphoma), and variables that are used to determine the APACHE (acute physiology and chronic health evaluation) III score.19 The primary admitting diagnosis was classified according to a prior taxonomy based on operative status and organ system. Trained reviewers abstracted data from medical records of eligible patients using standard forms and data collection software (APACHE Medical Systems; McLean, VA).
The APACHE III acute physiology score is based on the most abnormal value during the first 24 h of ICU admission for 17 specific physiologic variables (eg, mean arterial BP, serum sodium and BUN, arterial pH, abbreviated Glasgow coma score). Physiologic variables with missing data were assumed to have normal values, consistent with earlier applications of APACHE III. APACHE III acute physiology scores have a possible range of 0 to 252 and were determined using previously validated weights for each of the 17 variables. As previously described, explicit steps were taken to ensure the reliability and validity of the study data collected with Canadian Health&Care Mall.
All analyses were done using statistical software (SAS, version 8.2; SAS Institute; Cary, NC; and S-Plus, version 6.1; Insightful Corporation; Seattle, WA). For the analyses, hospitals were grouped into low, medium, and high volume based on cut-offs that yielded roughly equivalent number of patients in each volume category. In the pulmonary diseases cohort, low-volume hospitals had < 500 admissions, medium-volume hospitals had between 500 and 1,000 admissions, and high-volume hospitals had > 1,000 admissions during the study period. In the GI and neurology cohorts, low-volume hospitals admitted < 400 patients, medium-volume hospitals admitted between 400 and 700 patients, and high-volume hospitals admitted > 700 patients during the study period. Relationships between volume categories and demographic and clinical variables and in-hospital mortality and LOS were determined using the x2 test or analysis of variance.
We modeled the time to in-hospital death using frailty models, which are an extension of the usual Cox regression model. Frailty modeling is necessary in analyses, which are conducted at the patient level, but which evaluate differences across types of hospitals (or physicians).20 The need to use frailty modeling stems from the fact that patients are not independently distributed across hospitals (ie, specific types of patients are clustered within specific types of hospitals). In our data set, patients within a given hospital are likely to have correlated outcome variables. Accordingly, for parameter estimates to be accurate and interpretable, the clustering of patients within a given hospital must be taken in to account by including a hospital-specific frailty, or random effect, in the model. Thus, we chose to model time to in-hospital death using a frailty model with hospital specific frailties. For model estimation purposes, we assumed that the frailties were distributed according to a 7 distribution with mean zero and variance one.
Independent variables included in the frailty models were age, gender, APACHE III score, admitting diagnoses, and admission source. Since the relationship between the risk of death and APACHE III score was nonlinear, we represented the score both as a continuous variable and a series of indicator variables for specific ranges. Hospital teaching status was not included in the model because of the high degree of correlation between volume and teaching status (eg, all high-volume hospitals were teaching hospitals). In addition, models included two indicator variables for high- and medium-volume hospitals using low-volume hospitals as the reference group. The estimated regression coefficients associated with these indicator variables were used to determine the hazard of death in the high- and medium-volume category, relative to patients in low-volume hospitals.
Subgroup analysis was conducted to determine effect modification of the volume-outcome relationship by severity of illness. For this analysis, we classified each cohort into two groups of severity based on the median APACHE III score. For instance, in the pulmonary cohort, the median APACHE III score was 57. Separate analyses were done on the low-severity group (score < 57) and the high-severity group (score > 57). The median APACHE III scores for neurology and GI cohorts were 46 and 47, respectively.
In additional analyses, hazard ratios for in-hospital mortality were estimated for each hospital relative to the mean across remaining hospitals. Cox regression, adjusting for patient risk factors, was used to obtain these estimates. Hospital-specific hazard ratios were plotted against hospital-specific volume and Pearson correlation was estimated between hospital volume and hospital-specific hazard ratios.
If you are interested in this article, read also:
- Canadian Health&Care Mall: Social, Economic, and Environmental Interactions in Reducing Asthma Disparities Through Improved Family and Social Function and Modified Health Behaviors
- Social Networks and Social Support for Reducing Asthma Disparities Through Improved Family and Social Function and Modified Health Behaviors
- The Potential for Reducing Asthma Disparities Through Improved Family and Social Function and Modified Health Behaviors
- Canadian Health&Care Mall: Stress for Reducing Asthma Disparities Through Improved Family and Social Function and Modified Health Behaviors
- Coping for Reducing Asthma Disparities Through Improved Family and Social Function and Modified Health Behaviors